def parse_args():
parser = argparse.ArgumentParser('Get Test Result of Hateful Memes')
parser.add_argument('--cfg',
type=str,
help='path to answer net config yaml')
parser.add_argument('--ckpt',
type=str,
help='path to checkpoint of answer net')
parser.add_argument('--bs', type=int)
parser.add_argument('--gpus', type=int, nargs='+')
parser.add_argument('--model-dir',
type=str,
help='root path to store checkpoint')
parser.add_argument('--result-path',
type=str,
help='path to store test result file.')
parser.add_argument('--result-name', type=str)
parser.add_argument('--split', default='test')
args = parser.parse_args()
if args.cfg is not None:
update_config(args.cfg)
if args.bs is not None:
config.TEST.BATCH_IMAGES = args.bs
if args.gpus is not None:
config.GPUS = ','.join([str(gpu) for gpu in args.gpus])
if args.split is not None:
config.DATASET.TEST_IMAGE_SET = args.split
if args.model_dir is not None:
config.OUTPUT_PATH = os.path.join(args.model_dir, config.OUTPUT_PATH)
return args, config
def tune_vl_bert(config_path,
pl_ckpt_path,
num_samples=10,
num_epochs=10,
gpus_per_trial=2):
# scheduler = ASHAScheduler(
# metric="loss",
# mode="min",
# max_t=num_epochs,
# grace_period=1,
# reduction_factor=2)
reporter = CLIReporter(
parameter_columns=[
"lr", "weight_decay", "warmup_factor", "max_epoch", "batch_size"
],
metric_columns=["mean_accuracy", "training_iteration"])
param_config = {
"lr": 6.25e-7,
"weight_decay": tune.loguniform(1e-5, 1e-2),
"batch_size": 4,
"max_epoch": tune.choice([4, 6, 8, 10]),
"warmup_factor": tune.uniform(0, 1),
"warmup_steps": tune.uniform(100, 800),
}
scheduler = PopulationBasedTraining(time_attr="training_iteration",
metric="mean_accuracy",
mode="max",
perturbation_interval=2,
hyperparam_mutations={
"lr":
tune.loguniform(6.25e-6, 6.25e-8),
"batch_size": [1, 2, 3, 4],
})
update_config(config_path)
model_base_cfg = copy.deepcopy(config)
tune.run(partial(
_tune,
vl_bert_config=model_base_cfg,
pl_ckpt_path=pl_ckpt_path,
num_gpus=gpus_per_trial,
),
resources_per_trial={
"cpu": 4,
"gpu": gpus_per_trial,
},
config=param_config,
num_samples=num_samples,
scheduler=scheduler,
progress_reporter=reporter,
name="tune_vl_bert")
def parse_args():
parser = argparse.ArgumentParser('Train Cognition Network')
parser.add_argument('--cfg', type=str, help='path to config file')
parser.add_argument('--model-dir', type=str, help='root path to store checkpoint')
parser.add_argument('--log-dir', type=str, help='tensorboard log dir')
parser.add_argument('--dist', help='whether to use distributed training', default=False, action='store_true')
parser.add_argument('--slurm', help='whether this is a slurm job', default=False, action='store_true')
parser.add_argument('--do-test', help='whether to generate csv result on test set',
default=False, action='store_true')
parser.add_argument('--cudnn-off', help='disable cudnn', default=False, action='store_true')
# easy test pretrain model
parser.add_argument('--partial-pretrain', type=str)
parser.add_argument('--k-fold-i', type=int)
args = parser.parse_args()
if args.cfg is not None:
update_config(args.cfg)
if args.model_dir is not None:
config.OUTPUT_PATH = os.path.join(args.model_dir, config.OUTPUT_PATH)
if args.partial_pretrain is not None:
config.NETWORK.PARTIAL_PRETRAIN = args.partial_pretrain
if args.slurm:
proc_id = int(os.environ['SLURM_PROCID'])
ntasks = int(os.environ['SLURM_NTASKS'])
node_list = os.environ['SLURM_NODELIST']
num_gpus = torch.cuda.device_count()
addr = subprocess.getoutput(
'scontrol show hostname {} | head -n1'.format(node_list))
os.environ['MASTER_PORT'] = str(29500)
os.environ['MASTER_ADDR'] = addr
os.environ['WORLD_SIZE'] = str(ntasks)
os.environ['RANK'] = str(proc_id)
os.environ['LOCAL_RANK'] = str(proc_id % num_gpus)
return args, config
def train(config_path, pl_ckpt_path, **kwargs):
with logger.catch(reraise=True):
update_config(config_path)
cfg = copy.deepcopy(config)
_train(cfg, pl_ckpt_path, **kwargs)